The introduction of AgentBuild allows scientists to create LLM-based agents through a structured contract, enhancing workflow efficiency in scientific research. This method shifts the focus from individual judgment to a systematic rubric-driven approach, potentially improving the accuracy of data analysis. The application in Rietveld refinement of X-ray diffraction data demonstrates its practical utility and highlights remaining workflow limitations. As base models improve, AgentBuild can be easily re-tuned, preserving the scientist's original contributions.